Keeping company data safe in an AI knowledge assistant
Decide where it runs, connect only the sources you choose, make access mirror the access people already have, sign a data processing agreement naming every sub-processor and region, use AI providers on business terms that exclude training on your data, and agree in writing how data is returned or deleted at the end.
Where should the assistant run?
A knowledge assistant can run on the supplier’s cloud or on your own AWS, Google Cloud or Azure account. Your own account costs a little more and keeps the data inside your perimeter; for regulated data or strict clients it is usually the right call.
Which sources should you connect?
Start with the channels and folders that answer the questions people actually ask. Exclude HR, legal, finance and personal messages unless there is a reason to include them. Every source added later is a written decision, not a default.
Who should see what?
Nobody should get an answer from a source they could not open themselves. Where departments need different views, build role-based access on top of the assistant rather than one shared bot for everything.
What should the data processing agreement say?
Under UK GDPR you are the controller and the supplier is your processor. The data processing agreement should name the data, the purposes, every sub-processor including the AI provider, the storage regions and the retention period — and say that the AI provider’s business terms exclude training on your data.
Do people need to be told it is an AI?
Bots should say they are AI in their profile and first message; EU AI Act transparency rules apply to such systems from August 2026. Documents the assistant drafts should be reviewed by the person who sends them.
What happens to the data at the end?
Agree how data is returned or deleted when the project ends, how backups are purged, and that deletion is confirmed in writing. Our own approach is on the security page.
Questions
Do AI providers train on our data?
Under their business and API terms, OpenAI, Anthropic and Google exclude training on customer data. Check the terms of the plan you are on and name the provider in your DPA.
Can it run without sending data to an external model?
Yes, by running open-weight models on your own infrastructure. It costs more and answers can be weaker; it is worth it for some regulated data.
How we write these guides: figures about our own work come from before-and-after measurements on client projects (names withheld where under NDA); market figures come from the sources listed, with the date we checked them. Prices exclude VAT.
